Chapter 16

The Prefrontal Cortex of Agents: Reasoning Techniques

Interactive Graph (beta)

Toggle graph

Reasoning is the agent’s executive function — exposing intermediate steps, exploring alternatives, and refining outputs instead of answering in one shot.

Neuroscience Analogy

• PFC working memory: mental scratchpads for intermediate steps.
• Tree search: hippocampal “preplays” of possible paths.
• Error monitoring: ACC triggers refinements.
• Hybrid symbolic+neural: program‑aided reasoning.

Core Techniques

1. Chain‑of‑Thought (linear working memory).
2. Tree‑of‑Thought (branching exploration + selection).
3. Self‑Correction (meta‑reasoning loops).
4. Program‑Aided Reasoning (generate+execute code/symbols).
5. ReAct (thought–action–observation loops with tools).
6. Collective Reasoning (debate, chain/graph of agents).
7. MASS (optimize prompts and topology of agent collaboration).

Scaling Inference: Thinking Budget

More compute/time at inference → better reasoning, sometimes surpassing larger models that answer instantly.

Applications

Complex QA, math/code, debugging, strategic planning, medical/legal analysis, deep research with time budgets.

Conclusion

Reasoning techniques turn pattern‑matching into deliberate, transparent problem‑solving — the agent analogue of the PFC.